Cantina launches Apex Flash 1, an open‑weight model for focused software‑security investigations

Cantina Security, in partnership with the AI‑focused firm Yeta, announced the launch of Apex Flash 1, a reinforcement‑learning post‑training of the GLM‑5.3‑Flash architecture that has been specifically fine‑tuned for software‑security investigations. The company markets the model as a “focused worker” that can be summoned by a higher‑level supervisory agent to carry out concrete steps such as static and dynamic code analysis, tool invocation, exploit generation, and live verification of the exploit’s effect.

The underlying architecture comprises roughly 321.3 billion parameters, but thanks to a sparse‑activation mechanism only about 18 billion of those parameters are activated for each generated token. Although the open‑weight checkpoints are publicly released, the sheer size of the model still demands substantial memory bandwidth and GPU capacity, rendering it impractical for low‑end hardware unless users provision appropriate compute infrastructure.

Evaluation methodology and results

Cantina evaluated Apex Flash 1 on a curated benchmark consisting of 60 individual tasks derived from 20 previously unseen vulnerability cases. The benchmark was divided into three investigative views—guided white‑box, focused white‑box and focused black‑box—each executed inside isolated target environments that incorporated final‑state verifiers to confirm whether a generated exploit truly achieved its intended impact.

The model succeeded on 40 out of the 60 tasks, delivering a 66.7 % pass@1 rate on first‑draw attempts. Cantina reported an average computational cost of $2.38 per task, a figure that reflects the efficiency gains from sparse activation despite the model’s massive parameter count.

For context, Cantina’s internal comparison shows that the unmodified base GLM‑5.3‑Flash model solved 36 of the 60 tasks at a cost of $4.56 per task, while Anthropic’s Claude Opus 5.5 solved 43 of the 60 tasks but incurred a markedly higher cost of $74.68 per task. Cantina emphasizes that these numbers stem from its own evaluation pipeline and have not been independently audited.

Recommended usage pattern

The developers advise deploying Apex Flash 1 through the Codex agent harness, treating it as a specialized worker that operates under the guidance of a broader supervisory agent. This layered approach is intended to mitigate the risks associated with unsupervised, end‑to‑end security automation by ensuring that a human‑level controller can intervene, validate outputs, and enforce safety constraints.

The public evaluation focuses exclusively on text‑based security tasks. Apex Flash 1’s retained multimodal capabilities—such as processing images or video—were not part of the benchmark, and the reported results should not be extrapolated to scenarios that require visual or audio analysis.

Licensing and legal responsibilities

Apex Flash 1 is released under the permissive MIT licence, granting users broad freedom to use, modify, and redistribute the model and its associated code. Cantina, however, stresses that users remain fully accountable for obtaining proper legal authorisation, sandboxing the model, performing rigorous human review of any generated outputs, and preventing any offensive use beyond systems they are expressly permitted to test.

• Open‑weight checkpoint available on Hugging Face
• Sparse activation reduces per‑token compute
• Pass@1 of 66.7 % on 60 held‑out bug tasks
• Average cost of $2.38 per task
• Recommended integration via Codex agent harness

Two distinct checkpoints are distributed: the standard model and an experimental “abliterated” derivative that modifies the model’s refusal behaviour. Cantina clarifies that the published evaluation applies solely to the standard checkpoint; the abliterated version has not been benchmarked and should be treated as experimental.

Overall, Apex Flash 1 represents a notable step toward open‑source, high‑capacity models that can assist security researchers in automating repetitive investigative steps while still requiring human oversight for strategic decision‑making and ethical judgement.

For English‑speaking organisations, the practical impact is clear: teams can integrate Apex Flash 1 into existing security pipelines as a plug‑in worker, leveraging its ability to parse code, invoke analysis tools, and generate exploit snippets at a modest per‑task cost. By coupling the model with a supervisory agent and strict sandboxing, firms can accelerate vulnerability triage and proof‑of‑concept generation while maintaining compliance with legal and ethical standards.

Future outlook and community involvement

Cantina plans to extend the model’s capabilities through community‑driven fine‑tuning, inviting security researchers to contribute additional task sets, safety filters, and domain‑specific adapters. The company also signals an intention to publish a more comprehensive multimodal benchmark later this year, which would assess the model’s performance on image‑based code screenshots and video‑based debugging sessions.

In the coming months, Cantina expects to release tooling that simplifies the deployment of Apex Flash 1 on on‑premise clusters, as well as a lightweight inference wrapper that can run on GPU‑accelerated edge devices for red‑team exercises that require low latency.

Localisation : Paris, France – Les équipes de sécurité françaises peuvent dès à présent télécharger le checkpoint depuis Hugging Face, l’intégrer dans leurs chaînes CI/CD via le harness Codex, et profiter d’une solution open‑source qui respecte les exigences de souveraineté des données tout en offrant un coût d’exécution compétitif.